> How to set "write and read with QUORUM"? You set this through each thrift api call you are making through your java code. See http://wiki.apache.org/cassandra/API
> They are run physically separate hw (But same since they are VM) So they share disk. I think this can have an influence. As I understand it, cassandra can be fast on writes because its storage format is designed so that it needs to do very few random io, which results in fever seeks. And I think part of the good read performance comes from data on keys being distributed onto different nodes, which then means distributed reads on separate disks. When nodes share disk, I think a lot of this is negated.